Minimum supported PyTorch version?

[thomasjpfan]

Similar to #21 but for other array libraries. For example, PyTorch correction was is in 2.0, but not in previous versions. The following fails for PyTorch < 2.0:

array-api-compat/array_api_compat/torch/_aliases.py

Line 378 in e2203df

return torch.std(x, axis, correction=correction, keepdims=keepdims, **kwargs)

For reference, here are the torch.std docs for 1.13 and 2.0.

[asmeurer]

correction isn't documented, but it is present in 1.13.

>>> import torch
>>> torch.__version__
'1.13.0'
>>> torch.std(torch.arange(3, dtype=torch.float32), (0,), correction=2)
tensor(1.4142)

I've tested all the pytorch support against 1.13 so things should work there. I haven't tested against 2.0 yet, but don't expect too many breakages.

We should probably support at least 1.13 and 2.0 for now. Let me know if supporting anything earlier is important for you.

[rgommers]

correction was only recently updated to accept floats in addition to integers. Maybe that's the issue?

[thomasjpfan]

From the weekly call, correction must be an int to support 1.13.

[thomasjpfan]

As a minimum reproducer, the following fails on 1.13:

import torch
import array_api_compat 

x = torch.asarray([1.0, 2.0, 3.0])
xp = array_api_compat.get_namespace(x)

xp.std(x)
# TypeError: std() received an invalid combination of arguments - got (Tensor, tuple, correction=float), but expected one of:
# * (Tensor input, tuple of ints dim, *, int correction, bool keepdim, Tensor out)
# * (Tensor input, tuple of ints dim, bool unbiased, bool keepdim, *, Tensor out)
# * (Tensor input, bool unbiased)
# * (Tensor input, tuple of names dim, *, int correction, bool keepdim, Tensor out)
# * (Tensor input, tuple of names dim, bool unbiased, bool keepdim, *, Tensor out)

Setting correction to an int works, i.e. xp.std(x, correction=0).

[asmeurer]

The test suite xfail of this not working with nonintegral correction made me not realize this was broken with the default value. This is fixed in #25